v1.9.109-criteo7

Fix grace period bug (#33)

Following the change of task id / instance id format, task health status
was not properly tracked because bound to the same instance id after
successive kills for health check failure. This fix propose to address
the issue by tracking health status by task id to ensure it to cleaned
each time a task is terminated, whatever the reason.
Also fix similar bug with `killingInFlight` structure impacting anti-snowball.